Professional Documents
Culture Documents
1710 06957 PDF
1710 06957 PDF
VI. EXPERIMENTS AND RESULTS Comparing the mean response time of the SSQ strategy
to that of the Random strategy, we can see that it is worse
A. Comparing stratagies with up-to-date information. even if the stale period is only 4 seconds. But with 2
First of all we want to decide which strategy is better if seconds time between updates, SSQ strategy works better
we have up-to-date information every time the dispatcher than Random, but still much worse than USQ which is
makes a decision. In the Section IV Part C we showed that considered as an ideal strategy. Figure 3 shows the results of
the Shortest Queue load balancing strategy is optimal for the the experiments that prove these statements.
system we use. The experiments results are consistent with 6 Stale Shortest Queue (stale period 4 sec)
theoretical result. The Shortest Queue strategy works much
Random
better than Random and Round Robin strategies. The results
Mean Response Time (seconds)
8 2
6 1
4 0
0 10 2040 3050 60 70 80
2 Workload, %
Figure 3. Comparison of the mean response time of the Shortest Queue
balancing strategy with stale information to Random and Up-to-Date
0
Shortest Queue strategies. Two different stale periods are shown.
0 20 40 60 80
Workload, % C. Studing the HSQ strategy performance.
Figure 1. Increase of the mean response time depending on the system’s The purpose of developing a new load balancing strategy
workload for Random, Round Robin and USQ strategies.
is to decrease the mean response time of the system when
B. Using stale information in the dispatcher to make a stale information is used. Our experiments show that we
decision. have achieved this goal. Results, shown on the Figure 4,
It is practically impossible to get fresh information about indicate that HSQ works much better than SSQ with the
the system every time we need to transfer a customer to a same stale period. Also we can see that the delay increases
much slower when using the HSQ strategy instead of the
server. Thus, stale information is used in the real systems.
SSQ.
9 Shortest Queue with History (workload Thus, we can say that the value of the mean response
8
60%) time for the HSQ strategy in the case of a very long stale
Shortest Queue with History (workload period will be very close to those of the Round Robin
Mean Response Time (seconds)
80%)
7 strategy.
6
VII. CONCLUSION
5
In this paper, we studied four load balancing algorithms
4 using a simulation model consisting of one load balancer
3 and five servers. Results show that shortest queue with up-
to-date server load information gives the lowest delay as
2
expected. But if the load/state information is significantly
1 out of date then the shortest queue strategy performs worse
0
even than random load balancing. To improve stale
0 5 10 15 20 25 information based shortest queue, we have implemented a
Stale Period (seconds) new history based shortest queue load balancing strategy.
Figure 4. Mean response time for the Shortest Queue balancing strategy Although our new approach works worse than the optimal
with stale information and history depending on the stale period (the time strategy, the difference is not as large as for other strategies.
between information updates). Different workloads of the system are Even the state information is too old, our approach performs
presented. The results are compared to the regular Shortest Queue with close to round-robin strategy. The improved strategy can be
stale information. used in real system as it has low memory requirement and
Comparing the HSQ strategy to the USQ strategies, we we do not have to update the state information too
can see that although it works worse than the optimal frequently to get reduced delay compare to other load
strategy the difference is not as large as for other strategies. balancing strategy discussed in the paper.
The results of this comparison are shown on the Figure 5.
We also want to mention that even with extremely long time REFERENCES
between the updates the HSQ strategy performance is very [1] http://www.mesquite.com/
close to those of the Round Robin strategy. It can be [2] S. Zhou. A trace-driven simulation study of dynamic load balancing.
IEEE Transactions on Software Engineering, 14:1327–1341, 1988.
explained by careful examination of the strategy. If the stale [3] D. L. Eager, E. D. Lazokwska, and J. Zahorjan. Adaptive load sharing
period is very long and significant number of jobs arrives in homogeneous distributed systems. IEEE Transactions on Software
during it, then the HSQ strategy will work as follow: Engineering, 12:662–675, 1986.
[4] M. Dahlin. Interpreting Stale Load Information. IEEETransactions on
1) Equalizes the lengths of the servers queues. As it Parallel and Distributed Systems, 11(10),pp. 1033–1047, October 2000.
transfers the job to the server with the shortest queue and [5] M. Mitzenmacher. How useful is old information? In Principlesof
then increments the length of the queue, it will give jobs to Distributed Computing (PODC) 97, pages 83–91, 1997.
the servers with a less queue lengths until all the queue [6] Gupta, V., Balter, M. H., Sigman, K. and Whitt, W (2007). Analysis
of Join-the-Shortest-Queue Routing for Web Server Farms. Performance
lengths become equal. Evaluation. 64, 1062–1081.
2) The strategy starts to work in round robin fashion [7] W. Whitt. Deciding which queue to join: some counter examples.
Operations Research, 34:226-244, 1986.
since it has equlized the lengths of the servers queues. First [8] Milan E. Soklic (2002), "Simulation of Load Balancing Algorithms:
we will randomly choose one server from 5, then one from A comparative Study," SIGCSEBulletin, Vol.34, No.4, pp. 138-141, Dec.
the remaining 4, etc. Thus, we will schedule five next jobs [9] Mor Harchol-Balter, Task assignment with unknown duration,
to the five next servers and then will repeate this procedure Journal of the ACM (JACM), v.49 n.2, p.260-288, March 2002.
from the beginning. [10] W. Winston. Optimality of the shortest line discipline. Journal of
Applied Probability, 14:181-189, 1977
9 Shortest Queue with History (workload
60%)
8 Shortest Queue with History (workload
Mean Response Time (seconds)
80%)
7
6
5
4
3
2
1
0
0 105 15 20 25
Stale Period (seconds)
Figure 5. Comparison of the mean response time of the Shortest Queue
balancing strategy with stale information and history to Round Robin and
Up-to-Date Shortest Queue strategies. Two different stale periods are
shown.